Ghost Vessel — give your local agent a vessel

Bind your local LLM agent (the ghost) into an avatar body (the vessel). A monitor-resident, video-call-style avatar that fronts your personal AI agent (Hermes / OpenClaw, or anything) — a full replacement for the Telegram/messenger chat you keep open to talk to it, serving that agent's exact slash-command menu right in the avatar's window. Not a waifu toy: a real agent client that happens to have a face.

Your agent replies; the avatar performs it: fine-grained facial emotion beats, a persistent mood that sinks when you scold it and brightens when you praise it, idle states that breathe, blink, and sometimes rest their eyes — all from pre-rendered clips, so the GPU stays free at runtime (no live inference).

Demo

▶ Watch the full 45s demo with sound — emotion beats, the real Hermes slash menu, work loops, code cards, human-in-the-loop approval, and the floating monitor-resident windows.

The avatar above is Yeoreum, a photoreal preset sold separately. Cloning this repo gets you Haneul — a free 2D starter avatar that ships with the engine, so it runs the moment you clone it.

📦 Download for Windows — dependency-free (no Python), ships with both avatars. Pay-what-you-want lifetime key (package from $5, preset from $1) at ghostvessel.space.

📝 How it works — a writeup — the emotion-beat output contract, the no-runtime-GPU avatar, and hooking a real agent in as a connector.

How it works

you type ──► chat UI ──► bridge ──► relay connector ──► your agent (Hermes/OpenClaw)
                                                            │ reply (emotion-tagged)
video window ◄── performance player ◄── parser (3 planes) ◄─┘
  ▲ emotion beat clips + mood-based idle           │
  └── mood/affinity tracker (praise/scold, persistent state)

• 3-plane output: the reply splits into action (emotion beats) / dialogue (spoken via local TTS) / data (code & files → chat cards, never read aloud).
• Emotion engine: expression segments (valence/arousal-tagged), tag-less fallback (emoji/keyword inference), blink-aligned seamless idle loops, a head-pose "settle gate" so expressions reveal when the head is frontal.
• Mood & affinity: short-term mood decays toward a long-term relationship baseline. Keep scolding → it rests in a subdued idle. Persistent across restarts.
• Agent-agnostic: Hermes (relay connector contract, WS server) and OpenClaw (gateway WS client) adapters, plus a demo responder with zero setup. The agent is told only the emotions your avatar actually has, so partial avatars just work.
• The messenger's menu, in the avatar: type / and the chat pulls your agent's live command menu — the same one you'd see in Telegram (e.g. 52 commands for Hermes: /model, /new, /goal, /status, /compress, …), served per-agent (Hermes or OpenClaw). A real control surface, not a canned list; commands pass straight to the agent.
• Swappable voice: local (Qwen3-TTS with voice cloning, MeloTTS, Piper) or cloud (Edge — free & keyless, ElevenLabs, any OpenAI-compatible) — picked in settings, with an in-app installer that surfaces each engine's deps. Voice input via VAD + local STT.

Quickstart (Windows)

1. Prereqs: Python 3.11, an agent (optional — demo mode works without one).
2. Pick an avatar: the bundled Haneul works out of the box — echo gv-starter > presets/active.txt. (Bringing your own? Drop its folder into presets/ and name it there instead.)
3. Link your agent: python bridge/setup_connector.py (picks Hermes/OpenClaw, Windows/WSL2, writes the connection + injects the avatar output contract into the agent's prompt, scoped to this channel).
4. Run the stack — TTS (:8899), bridge (:8900), connector (:8901), and a static server for the player (python -m http.server 8777). Then either open http://127.0.0.1:8777/player/index.html in a browser, or run the desktop shell (cd src-tauri && cargo tauri dev) for a frameless, always-on-top avatar window. (Each component is a small local server — wire up a launch script for your own setup.)

Avatars (presets)

An avatar is a pure-data bundle (clips + persona + theme + voice + emotion map) — no code runs, so installing one is safe. Presets are folder-mapped: drop a folder of clips named by convention (happy.mp4, angry.mp4, idle.mp4, …) into presets/ and the engine maps them automatically — the folder name is the avatar's name.

A free starter avatar ships with the engine, so cloning the repo gives you something that actually runs: Haneul (presets/gv-starter/) — 2D anime, 6 emotion beats, a blink-aligned idle, mood-based resting idles, and a talking loop.

She was built end to end with free local tools — Animagine XL for the neutral source still, HunyuanVideo 1.5 (480p i2v, step-distilled) for the clips, then the tools/ scripts below to loop and validate them. That's the same path this README asks you to walk, on one 12GB card, with no paid service anywhere in it — so the method is demonstrated rather than just described.

The clip on the right is her actual happy / surprise / angry / downcast segments, straight out of presets/gv-starter/avatar/segments/.

The demo at the top of this README is Yeoreum, a photoreal avatar sold separately — different tier, same engine. Nothing in the engine treats them differently.

Any art style works. The engine plays video clips and doesn't care what produced the pixels — photoreal, 2D anime, a 3D render, pixel art, an abstract shape. You can even take an existing Live2D or VRM model and pre-render its expressions into clips; a rig-based shell can't do the reverse.

Make your own. The bundle/filename spec is in docs/PRESET_FORMAT.md, and tools/ ships the authoring toolchain so you don't have to solve the fiddly parts yourself:

1. One neutral, front-facing still of your character.

2. Animate it into expression clips with an image-to-video model. The engine only needs MP4s, so any i2v model works — but the choice matters more than you'd think:

   • HunyuanVideo 1.5 (480p i2v, step-distilled) — free, local, and what the starter avatar was built with. Runs on a 12GB card at ~170s per 5-second clip. It's the smallest model we found that actually renders subtle expressions — smaller video models (we tried Wan 2.2 5B) will happily do an open-mouthed grin but render "concerned" and "angry" as a blank neutral face.
   • Cloud i2v (Gemini/Veo, Higgsfield, …) also works and needs no GPU — the paid Yeoreum avatar's clips were made this way. Costs credits.
   • LivePortrait is a lighter local option that retargets driving videos onto your still, but it copies a real face's motion rather than acting your prompt, and it can get uncanny on stylised art.

   Prompt the whole arc — "calm neutral face at first, then , then returns to the neutral calm expression" + "static camera, fixed framing, only her face changes" — and describe expressions physically ("eyebrows slant down into a hard V"), not by label. Don't put "extreme expression" in the negative prompt; it flattens every expression you asked for.

3. Cut a multi-emotion take into segments — only if your model packs several expressions into one take (cloud i2v often does). tools/cut_emotions.py --video take.mp4 --strip prints a contact sheet so you can spot the neutral valleys, then --cuts 3.4,6.7 --emotions shy,happy,surprise slices and web-encodes them. One-emotion-per-clip? Skip this step.

4. Build a seamless idle loop — tools/build_idle_loop.py --video idle.mp4 --out presets/<id>/avatar. Seamless looping is the part that actually takes effort, so this does it for you: it runs MediaPipe over the take, finds the blink minima, and picks the blink→blink window that maximizes eyes-open time and pose match — so the loop's seam lands on a closed eye and is invisible. Falls back to pingpong when no clean blink pair exists, and records the head-frontal "settle times" the player uses to time expression reveals.

5. Check it — tools/validate_preset.py presets/<id> gates on structure, asset integrity, and emotion coverage. Exit 0 = shippable.

Drop the folder into presets/ and you're done. Keep it SFW; own your likeness rights.

Two traps worth knowing, since presets/gv-starter/ is a worked example of both:

• Your i2v model will probably ignore "then returns to neutral" and just hold the expression to the last frame — which makes the cut back to idle jump. Trim at the peak and mirror the clip; the expression forming, reversed, reads as the expression relaxing.
• emotion_map.json needs an axis ({emotion: {valence, arousal}}). It's what the mood tracker reads — without it every emotion scores valence 0 and the whole mood/affinity system silently does nothing. Same for reactions: the defaults name emotions (excited, wince, frown_subtle) your preset may not ship, so remap them to yours.

Prefer to skip the production step? Get the demo avatar (Yeoreum) — photoreal, 16 emotions, pay-what-you-want — or commission a custom one (open an issue or ask at ghostvessel.space).

License

Everything in this repository is MIT (see LICENSE) — the engine, the tools/ scripts, the docs, and the bundled Haneul starter avatar. Clone it and it's yours, batteries included.

The one exception isn't in the repo: Yeoreum, the photoreal avatar sold at ghostvessel.space, is a commissioned likeness and ships under its own personal-use license — as does any third-party preset you install from elsewhere.

Open-sourced in donation-ware spirit — issues / PRs welcome. 🛠️